Delete S1/12/.ipynb_checkpoints/tripletmarginloss_cuda-checkpoint.py

S1/12/.ipynb_checkpoints/tripletmarginloss_cuda-checkpoint.py

@@ -1,108 +0,0 @@
-# example_cudacode.py
-
-import torch
-from torch.utils.cpp_extension import load_inline
-
-# Triplet Margin Loss的自定义CUDA实现
-triplet_loss_source = """
-#include <torch/extension.h>
-#include <cuda_runtime.h>
-#include <cmath>
-
-// 定义共享内存和块的大小
-#define BLOCK_SIZE 256
-
-// 计算两个向量之间的L2距离的平方
-__device__ float squared_l2_distance(const float* v1, const float* v2, int dim, float* sdata) {
-    float my_sum = 0.0f;
-    // 每个线程计算一部分元素的平方差之和
-    for (int i = threadIdx.x; i < dim; i += blockDim.x) {
-        float diff = v1[i] - v2[i];
-        my_sum += diff * diff;
-    }
-    sdata[threadIdx.x] = my_sum;
-    __syncthreads(); // 确保所有线程都完成了它们的初始求和
-
-    // 使用共享内存执行并行规约
-    for (unsigned int s = blockDim.x / 2; s > 0; s >>= 1) {
-        if (threadIdx.x < s) {
-            sdata[threadIdx.x] += sdata[threadIdx.x + s];
-        }
-        __syncthreads();
-    }
-
-    // 最终的平方和在sdata[0]中
-    return sdata[0];
-}
-
-__global__ void triplet_loss_kernel(
-    const float* anchor, const float* positive, const float* negative,
-    float* loss, int batch_size, int dim, float margin) {
-
-    int batch_idx = blockIdx.x;
-    if (batch_idx >= batch_size) return;
-
-    __shared__ float sdata[BLOCK_SIZE];
-
-    // 计算当前三元组的向量指针
-    const float* anchor_ptr = anchor + batch_idx * dim;
-    const float* positive_ptr = positive + batch_idx * dim;
-    const float* negative_ptr = negative + batch_idx * dim;
-
-    // 计算正对和负对的距离
-    float dist_pos_sq = squared_l2_distance(anchor_ptr, positive_ptr, dim, sdata);
-    float dist_neg_sq = squared_l2_distance(anchor_ptr, negative_ptr, dim, sdata);
-
-    // 只有块中的第一个线程执行最终的计算和写入操作
-    if (threadIdx.x == 0) {
-        float dist_pos = sqrtf(dist_pos_sq);
-        float dist_neg = sqrtf(dist_neg_sq);
-        float loss_val = dist_pos - dist_neg + margin;
- 
-        loss[batch_idx] = fmaxf(0.0f, loss_val);
-    }
-}
-
-torch::Tensor triplet_loss_cuda(
-    torch::Tensor anchor, torch::Tensor positive, torch::Tensor negative, float margin) {
-    
-    int batch_size = anchor.size(0);
-    int dim = anchor.size(1);
-    
-    auto options = torch::TensorOptions().device(anchor.device()).dtype(anchor.dtype());
-    auto loss = torch::empty({batch_size}, options);
-
-    dim3 grid(batch_size);
-
-    dim3 block(BLOCK_SIZE);
-
-    triplet_loss_kernel<<<grid, block>>>(
-        anchor.data_ptr<float>(), positive.data_ptr<float>(), negative.data_ptr<float>(),
-        loss.data_ptr<float>(), batch_size, dim, margin);
-    
-    return loss.sum()/batch_size;
-}
-"""
-
-triplet_loss_cpp_source = """
-torch::Tensor triplet_loss_cuda(
-    torch::Tensor anchor, torch::Tensor positive, torch::Tensor negative, float margin);
-"""
-
-# 编译内联CUDA代码
-triplet_loss_module = load_inline(
-    name="triplet_loss_module",
-    cpp_sources=triplet_loss_cpp_source,
-    cuda_sources=triplet_loss_source,
-    functions=["triplet_loss_cuda"],
-    verbose=True
-)
-
-class ModelNew(torch.nn.Module):
-    def __init__(self, margin: float = 1.0):
-        super(ModelNew, self).__init__()
-        self.margin = margin
-        self.triplet_loss = triplet_loss_module
-
-    def forward(self, anchor, positive, negative):
-        return self.triplet_loss.triplet_loss_cuda(anchor, positive, negative, self.margin)